You are on page 1of 37


32, 81 1-847 (1991)



Applied Mechanics Division I l l , Sandia National Laboratories, Albuquerque, New Mexico 87185, U.S.A.

Civil Engineering Department, Brigham Young University, Provo, Utah 84602, U.S.A.

This paper presents a new mesh generation technique,paving, which meshes arbitrary 2-D geometries with
an all-quadrilateralmesh. Paving allows varying element size distributions on the boundary as well as the
interior of a region. The generated mesh is well formed (i.e. nearly square elements, elements perpendicular
to boundaries, etc.) and geometrically pleasing (i.e. mesh contours tend to follow geometric contours of the
boundary). In this paper we describe the theory behind this algorithmic/heuristictechnique, evaluate the
performance of the approach and present examples of automatically generated meshes.

The finite element method is a fundamental numerical analysis technique in wide-spread use in
the engineering community. It has long been recognized that this technique is a keystone in the
design and understanding of complicated mechanisms and processes. However, one of the biggest
obstacles that must be overcome for this method to be used is the discretization of a general
geometry of the problem into a valid finite element mesh. Not only is the mesh generation task
tedious and error prone, but the accuracy and cost of the analysis direcly depends on the size,
shape and number of elements in the mesh.
Since automation of mesh generation has a potentially large payoff, extensive research
continues in this area. Most of the research focuses on techniques such as the 2-D quadtree and
3-D ~ctree,'-~ triangulation4, or substructuring.6 In general, these techniques generate meshes
using triangular elements that are less accurate and less versatile than quadrilateral elements.
Some of these techniques have been extended to produce all-quadrilateral meshes. However, most
commercial mesh generation codes still employ parameter space mapping7, to generate an all-
quadrilateral mesh.' This technique is not automatic, in that the analyst must decompose the
geometry into regions that map well into parametric space. Even though the mapping technique
is laborious and expertise-intensive, many skilled analysts prefer it over the other techniques for
the following reasons:
Boundary sensitiue. Mesh contours closely follow the contours of the boundary. This
characteristic is of particular importance since well shaped elements are usually desirable
near the boundary.
*This work was performed at Sandia National Laboratories and supported by the U.S. Department of Energy under
contract DE-AC04-76DP00789

0029-598 1/9 1/1208 1 1-37$18.50 Received 15 April 1990

0 1991 by John Wiley & Sons, Ltd. Revised 15 May 1991

0 Orientation insensitive. Rotating or translating a given geometry does not change the
resulting mesh topology. A mesh generated in a transformed geometry is equivalent to the
original mesh transformed.
0 Few irregular nodes. An interior mesh node is considered irregular if more or less than four
elements are connected to it. This is a critical mesh topology feature because the number of
elements sharing a node controls the final shape of the elements, even after smoothing.1°
Thus, a mesh with few irregular nodes, especially near the boundary where element shape is
critical, is often preferred.
Noting these positive characteristics of mapping, we previously attempted to automate
quadrilateral meshing by using an expert system, AMEKS," to decompose a geometry into
regions for mapping (similar to the approach of experienced analysts.) However, AMEKS
suffered from a lack of robustness, long computer run times, and both software and hardware
incompatibility with most engineering environments. These concerns led to the development of
the paving technique.
The paving technique is designed to automatically generate an all-quadrilateral mesh with
qualities similar to those of the mapped mesh technique but without decomposition of the
geometry. This is accomplished by layering or paving the geometry with rows of quadrilateral
elements from the boundary@)toward the interior. The technique is relatively fast, robust and
capable of handling mesh transitions from small to large elements much more elegantly than
mapping techniques. Eventually this technique may be combined with automated decomposition
and/or adaptive finite element error correction techniques to produce an optimal mesh for a finite
element problem. The paving technique has been incorporated into an existing quadrilateral
mesh generation program, FASTQ,l2 and is being used in a production environment.
In this paper, we first present the theory of paving. Since paving is accomplished with an
interdependent series of steps, the theory is explained with an introductory overview of each step,
a general schematic showing how the steps interact, and finally a thorough explanation of each
step in turn. Following the theory presentation, we next evaluate the performance of the
technique both qualitatively with several examples and quantitatively using several evaluation
criteria. Finally, we end with a summary of our conclusions.

The paving technique is based on iteratively layering or paving rows of elements to the interior of
a region's boundary@).As shown in the brief sequence in Figure 1, these rows eventually fill the
region from the boundary(s) inward. As rows begin to overlap or coincide at the interior of the
geometry, they are carefully connected together to form a valid quadrilateral mesh. In this section
we first define the terminology and the input requirements for paving, second, give an overview of
each of the paving steps, and third, briefly describe how these steps interact (i.e. the control
sequence) during the paving process. From there, each of the steps is elaborated on, including (in
order of discussion) row choice, closure checks, row generation, smoothing, seaming, row
adjustment, perimeter intersection handling and cleanup of the completed mesh.

2.1. Terminology and input requirements

Paving begins with the input of one or more ordered, closed loops of connected nodes that, as
shown in Figure 2, form the boundary of the region to be meshed. These initial loops will be
referred to as the permanent boundary of the region. The connectivity and location of nodes on

Figure 1. Example of a simple paving sequence



t t
Figure 2. Example of the permanent boundary input for paving

the permanent boundary are not allowed to change during paving. This ensures mesh compatibil-
ity with adjacent regions.
Permanent boundaries are categorized as either exterior or interior boundaries. There is always
only one exterior permanent boundary for a region. It must be a loop of nodes that is non-
intersecting and completely encloses the region to be meshed. It is ordered counterclockwise.
Interior permanent boundaries define voids or holes in the region being meshed. They are always
ordered clockwise. There may be any number of interior permanent boundaries but each must be
completely enclosed within the exterior permanent boundary. No permanent boundary may
intersect any other permanent boundary.
During the mesh generation process, the paving technique always operates on boundaries of
connected nodes referred to as paving boundaries. Paving boundaries are transient in nature and
change as the mesh is generated. As shown in Figure 3, the paving boundaries (highlighted in bold
lines) always bound the existing mesh. Initially each paving boundary is identical to a permanent
boundary. Elements that are added to the mesh eventually separate the paving boundary(s) from
their respective permanent boundary@).
As with permanent boundaries, paving boundaries are categorized as either exterior or interior
paving boundaries. An exterior paving boundary is paved counterclockwise and progresses
inward from the exterior boundary. It is always ordered counterclockwise. Although initially
there is only one exterior paving boundary in a region, during paving any number may form.
Likewise, an interior paving boundary is paved clockwise and progresses outward from an
interior permanent boundary.





Figure 3. Example of interior and exterior paving boundaries

The nodes in the mesh are also categorized as shown in Figure 3. A paving node is defined as
any node on a paving boundary. A fixed node is defined as any node on a permanent boundary. A
floating node is defined as any node not on a permanent boundary. Thus, a paving boundary may
consist of fixed nodes, floating nodes, or a combination of fixed and floating nodes. Every paving
boundary node also has an interior angle which is the angle between a line connecting the node to
the preceding node and one connecting it to the next node in the paving boundary, as shown in
Figure 3.
Each paving boundary must always contain an even number of nodes. This is a necessary
condition when generating an all-quadrilateral mesh. This condition is enforced by controlling
the operations performed during paving.
The size of the elements in the mesh is determined by the spacing of nodes on the paving
boundary as it propagates. The spacing on a paving boundary is, of course, initially defined by the
fixed node spacings on the corresponding exterior boundary. Every attempt is made to maintain
this spacing or size as the boundary progresses. Thus, the user can control the element size by
modifying the spacing of initial nodes on a permanent boundary. This spacing need not be
uniform, and usually is not, as it is often desirable to mesh a region with a gradation of element

2.2. Overview of paving steps

The propagation of a paving boundary involves a number of operations which must be tightly
controlled to ensure mesh validity and quality. Each of the paving operations is explained in
detail in subsequent sections but may be summarized as follows:
Row choice. The beginning and ending node of the next sequence or row of elements to be
added is found.
Closure check. A check is made to make sure that more than six nodes remain in the paving
boundary. Specific closure techniques are used to conclude meshing for paving boundaries
of six or fewer nodes.
Row generation. The next row of elements identified in the row choice step is incrementally
added to the boundary.

4. Smooth. Floating nodes are adjusted to improve mesh quality and boundary smoothness.
5. Seam. Small interior angles in the paving boundary are seamed or closed by connecting
opposing elements.
6. Row adjustment. The new row is adjusted by placing tucks or wedges into the row to correct
for elements becoming too small or too large.
7. Intersection. The paving boundary is checked for intersections with itself or with other
paving boundaries. Intersections are connected to form new, often separate, paving bound-
8. Cleanup. The completed mesh is adjusted where element deletion and/or addition improves
the overall mesh quality.

2.3. Control sequence

The paving process is an iterative method which uses the steps outlined above. The control of
these steps is critical to the formation of a high quality mesh. The control sequence used for
paving is diagrammed in Figure 4.
As can be seen in Figure4, these steps sometimes interact in complex ways. For instance,
during a row generation, it may be necessary to suspend creation of the row at critical points,
smooth, look for seams and intersections, and then continue with the row generation if
appropriate. Also, whenever an intersection occurs, the mesh is smoothed and the boundary is
seamed on both sides before continuing to the choice of the next row. Iteration continues until all
paving boundaries close, thus completing the mesh.

2.4. Row choice

The first step in the paving technique is determining the location for the next row of elements. It
is critical that elements be placed by rows and that those rows be chosen carefully in order to
maintain boundary sensitivity, minimize the number of irregular nodes and control the placement

perform ROW CHOICE
ADD ROW portion
SMOOTH row portion
SEAM boundary
i f INTERSECTION occurs then
connect overlaps
SEAM boundary
end i f
until complete row i s added
i f INTERSECTION occurs then
connect overlaps
SEAM boundary
end i f
until CLOSURE CHECK i s p o s i t i v e

Figure 4. Paving control sequence


of irregular nodes. A row is defined by an appropriate beginning and ending paving node. This
involves a classification of the paving nodes on the paving boundary, a careful watch for paving
boundaries which form simple shapes (e.g. rectangle) or what we have termed primitive regions
and the consecutive insertion of new rows. Each of these will be described in turn.

Node classijication. To begin the row choice process, each paving node on the current paving
boundary is first given an angle status and finally a node classification. The angle status is a
categorization that is based solely on the size of the node’s interior angle, a (see Figure 3), and the
setting of angle tolerances, ol,.. . , 0 6 .The angle tolerances represent an allowable deviation
from an expected angular value. The angle status categories are:
Row end (a ,< n/2 gl).A row end node is a node at which a row terminates. Only one new
element will be inserted at a row end node, as shown in Figure 5.
Ambiguous row endlside ( n / 2 rsl < a < n - 0 2 ) .An ambiguous row end/side node may
function as either a row end or a row side.
Row side (n - rs2 < a < n 0 3 ) .A row side node is a node along which the row continues
normally. Two new elements are attached to each row side node, as shown in Figure 5.
Ambiguous row sidelcorner (n o3 < a < 3n/2 - a4). An ambiguous row sidefcorner node
may function as either a row side or a row corner.
Row corner ( 3 n / 2 - g4 < a < 3 n / 2 + rs5).A row corner node is a node at which the row
turns a logical corner. Three new elements are attached to each row corner, node as shown in
Figure 5.
Ambiguous row corner/reversal(3n/2 + ( T ~< a d 2n - 66). An ambiguous row cornerjrev-
ersal node may function as either a row corner or a row reversal.
Row reversal (a > 2n - 06). A row reversal node is a node at which the row turns two
sequential logical corners, or reverses its direction. Four new elements are attached to each
row reversal node, as shown in Figure 5.
These classifications are summarized in Figure 6. All nodes are eventually classified as either a
row end, side, corner, or reversal node. As can be seen in Figure 6, the chances of a paving node
falling into one of the ambiguous angle status categories are high. These ambiguities are resolved
by looking for simple shapes or primitives, or by avoiding irregular node generation as explained
Most paving boundaries contain two row end nodes where the next row can start and end.
However, there are three special cases of paving boundaries that do not contain two row end

Row Si

\‘Row End Nodes

Figure 5. Example boundary node classifications for a new boundary row
, .
# 4 ,
* Ends
d 8 icornersl

4 ! Sides c
i - Rebersols




Figure 6. Angle status categories based on paving node’s interior angle



Figure 7. Handling of special case paving boundaries for row choice

nodes and must be handled differently, as shown in Figure 7. The elements shown in Figure 7
indicate how the first row would be inserted in these special case geometries. Case 1 is a paving
boundary with only row side nodes (e.g. a circle). In this case the node with the smallest angle is
chosen as both the beginning and ending node in the new row to be generated. Case 2 is a paving
boundary with only one row end node and any number of row side, corner, or reversal nodes (e.g.
a heart shape). In this case the row end node becomes both the start and finish for the new row.
Case 3 is a paving boundary with no row end nodes but with some row corners and/or row
reversals (e.g. a club shape). In this case the row corner or row reversal node with the smallest
interior angle is chosen as the row start and finish node.

Primitive classifcation. The second factor considered when determining the next row to be
paved is the formation of simple or primitive shapesi3* Determination of primitive shapes
involves the formation of a more global perspective of the current paving boundary. If primitive
shapes are identified from this global perspective, the placement of new rows can be adjusted so as
to complete the paving of the primitive with a minimum of irregular nodes. The primitives
currently used in paving are the rectangle, the triangle and the semicircle. The search for
primitives is based on possible combinations of all row ends and row sides in the paving
boundary. Since a primitive, by definition, does not contain any row corners or reversals,
boundaries that contain row corner or reversal nodes are not evaluated as primitives.
Primitive classification begins with a combinatorial search of all possible row end and row side
node interpretations. All nodes which have been categorized as ambiguous row end/side nodes
are evaluated as being either a corner or a side for each of the possible interpretations. In order to

*Primitive classifications are considered only when dealing with an exterior paving boundary

preserve the possibility of a primitive shape, any ambiguous row side/corner nodes are always
interpreted as sides. Although it is possible that a given region may have a huge number of
interpretations, in practice the maximum number is around 10 and the average is 1 or 2.
Each combination of boundary interpretation is evaluated for quality and validity as a
primitive. The quality is based on the formation of irregular nodes along the existing boundary,
the deviation of the angles from perfect end or side nodes and the number of irregular nodes to be
expected within the interior of the resulting primitive as it is meshed. The validity depends on the
requirements of each primitive, as documented in Reference 13. The primitive interpretation that
is valid and has the highest quality is chosen as the correct interpretation.
For example, the paving boundary shown in Figure 8(a) contains two row end nodes, nodes 1
and 3, two ambiguous row end/side nodes, nodes 2 and 4, and the rest row side nodes. With this
set of nodes, there are four possible interpretations:
1. Nodes 2 and 4 may both be classified as row side nodes. In this case a valid semicircle
primitive (two end nodes) would be formed. Its quality would be affected by the fact that the
interior angles of side nodes 2 and 4 are significantly less than 71, the fact that node 4 would
become an irregular node, and the fact that a semicircle forms two additional irregular
nodes, as shown in Figure 8(b).
2. Node 2 may be classified as a row end node, and node 4 as a row side node. In this case an
invalid triangle primitive (three end nodes) is formed.
3. Node 2 may be classified as a row side node, and node 4 as a row end node. Again, an invalid
triangle primitive (three end nodes) is formed.
4. Nodes 2 and 4 may both be classified as row end nodes. In this case a valid rectangle
primitive (four end nodes) would be formed. Its quality would be affected by the fact that the
interior angles of end nodes 2 and 4 are significantly greater than 71/2 and the fact that node
2 would be an irregular node. However, the rectangle does not form any additional irregular
nodes, as shown in Figure 8(c), and in this case would be the best interpretation.
If an acceptable primitive evaluation exists, that evaluation is used in determining the
placement of the next row. Node classifications are set to those used in the primitive interpreta-

(a) O r i g i n a l P a v i n g Boundary

Ibl Paved As R S e m i c i r c l e Icl Paved As R RectangLe

Figure 8. Example of multiple primitive region interpretationsfor paving


tion and the row end nodes for the next row chosen so as not to invalidate the primitive while
maintaining sequential ordering as described in Section 2.4.
This evaluation technique is only a limited global view of the boundary since it does not
contain information as to the overall shape, the position and orientation of various boundaries
with respect to each other, the proximity of possible overlaps, etc. Because of the limitations in
this simple evaluation technique, boundary shapes are often misinterpreted. However, since the
interpretation mechanism is repeated as each row is added, the evaluations are self-correcting.
For example, the paving boundary in Figure 9(a) is initially misinterpreted as a rectangle.
Adding the first row of elements creates a self-intersectingpaving boundary that is subsequently
corrected, as shown in Figure 9(b). This forms two exterior paving boundaries which are both
correctly interpreted as triangle primitives. These paving boundaries are then filled with only one
irregular node each, as shown in the completed mesh in Figure 9(c). As shown with this example,
incorrect interpretations of complex shapes as being primitive do not adversely affect the
deposition of rows of elements. On the other hand, correct interpretations of simple shapes often
minimize the number of irregular nodes.
When an acceptable primitive evaluation does not exist, all ambiguous nodes are classified
based on angle size and the possibility of generating an irregular node. For example, if classifying
an ambiguous side/corner node as a side node would form an irregular node while a corner
classification would not, the corner classification is allowed a greater angular tolerance.

Consecutive ordering of rows. The final concern in the choosing of rows is that the paving
boundary be propagated successively inward away from all permanent boundaries. This is
accomplished by preferring new rows that are sequential to (i.e. topologically next to) the last row
installed in the paving boundary. Rows are initially installed around exterior paving boundaries
sequentially counterclockwise. After each complete traversal of the exterior paving boundary,
each interior paving boundary is paved with one complete traversal sequentially clockwise.

( a ) Incorrect Rectonqle Desiqnation

(bl C o r r e c t T r i a n g t e D e s i g n o t i o n s

Ic) F m i s h e d Mesh

Figure 9. Example of incorrect primitive region interpretation


RECTANGLE :..........................................
.......................................... TRIANGLE i
j 2-1-2-1 j 2-2-2 i 3-3 i 0

.............................................................................................................................. ~ ...................

Figure 10. Closure of boundaries with six nodes

Although rows may not form true rings inside the boundary, the sequential ordering of row
placement generally forces boundaries to close toward the interior of the region. This tends to
place irregular nodes interior to the geometry, away from the more critical boundary elements.

2.5. Closure check

Before any new rows of elements are added to the mesh, a check is made to see if the paving
boundary contains six or less nodes. If the boundary contains zero nodes, paving of the boundary
has been completed. If the boundary contains two nodes, paving can be completed with a simple
seam, as explained in Section 2.8. If the boundary contains four nodes, one additional quadrilate-
ral element is added to close the boundary. Closure for boundaries containing six nodes is based
on the configuration of those nodes and may close to form two, three, or four new elements, as
shown in Figure 10. The numbers above each case in the figure indicate how many element edges
are on each side of the shape (e.g. in a 3-2-1 triangle, the triangle's first side is formed from the
combination of three element sides, the second from two element sides, and the third from one
element side). Some of the elements shown in Figure 10 are poor quadrilateral elements.
However, because closures usually form in the centre of the mesh, subsequent smooths and
cleanup almost always improve or eliminate the problem elements.

2.6. Row generation

Once the limits of the new row have been chosen, actual generation can begin. Row generation
is accomplished by projecting new nodes away from each sequential boundary node at an
appropriate angle and distance. This projection is based on the node classification (Section 2.4)
and the distance between immediately preceding and subsequent boundary nodes. Elements and
nodes are always generated sequentially along the row. Projections from row side, corner and
reversal nodes, as well as from one additional special case, are described in the present section.?
Finally, row termination is discussed.
+Row end nodes form the basis for either beginning or ending a row and as such they are not projected

Row side node projection. The projection from a row side node is shown in Figure 11. Let the
existing side node be N , . The new node, N j , is placed at the tip of a vector V whose origin is at Ni.
V is oriented to bisect the interior angle, a, at N , . Let d , be the distance from N , to Ni- and d , the
distance from N , to N i +. Then the length of the vector V, I V I, is given by

The sine term in equation (1) tends to smooth out otherwise bumpy rows. One new element
defined by the nodes N , , N , - , , N j - l and N j is added to the mesh.
As can be seen in Figure 11, the projection from a row side node ( N , ) usually transforms the
node into a row end node, effectively moving the row end forward one step.

Row corner node projection. The projection from a row corner node is shown in Figure 12. Let
the existing corner node again be N , . Three new nodes, N j , N k and N , , are placed at the tips of

Figure 11. Projection of a new node from a row side node

Figure 12. Projection of new nodes from a row comer node


vectors V j ,v k and V , respectively. All three vectors originate at N , . V j ,v k and V, are oriented at
respectively the clockwise 1/3,1/2 and 2/3 bisections of the interior angle of N i . As before, let d , be
the distance from N i to N i - and d2 the distance from N i to N i +,. The magnitudes of the vectors
are given by

Two new elements are added to the mesh: the first defined by nodes N i , N i - N j - and N j , and
the second defined by nodes N i , N j , N k and N , .
As alluded to in the overview in Section 2.2, after the first element has been added at a row
corner, the operations of smoothing, seaming and intersection checking are performed. This is
necessary because in some cases a seam of the completed portion of the row may be appropriate.
Such a seam may place node N i interior to the mesh, preventing continuation of the row.
Likewise, an intersection of the row portion may be found which forms new boundaries and may
either place N i interior to the mesh or on a newly formed boundary. Again this would pre-empt
continuation of the row. In such a situation, control transfers back to the row choice procedure
described in Section 2.4 for determination of a new row to be paved.
As can be seen in Figure 12, the projection from a row corner node ( N i )not only propagates the
row end around the corner, but introduces another new row corner node (Nk) into the paving

Row reversal node projection. The projection from a row reversal node is shown in Figure 13.
Let the existing corner node again be N i . Five new nodes, N j , N k , N , , N , and N , , are placed at the
tips of vectors V j ,V , , V , , V , and V , respectively. All five vectors originate at N i . Here V j ,v k , V I ,
V , and V , are oriented at respectively the clockwise 114, 3/73, 1/2, 5/8 and 314 bisections of the
interior angle of N , . As before, let d , be the distance from N i to N i - l and d2 the distance from N ,

/ N"
Figure 13. Projection of new nodes from a row reversal node

to N i + l respectively. The magnitudes of the vectors are given by

Three new elements are added to the mesh: the first defined by nodes N,, Ni- 1, Njdl and N j ,
the second defined by nodes N,, Nj, Nkand N,, and the third element defined by nodes N,, N,, N,
and N,.
Again, as mentioned for a row corner node projection, after the first and again after the second
element has been added a smooth, seam and intersection check is performed to insure that
continuation of the row is beneficial.
As can be seen in Figure 13, a row reversal node projection usually produces two new row
corner nodes. This means that a row reversal node can occur only on the permanent boundary as
the first row immediately buries the very large interior angle.

All side node row projection. Of the special cases mentioned in Section 2.4, only one, the case of
a row with only row side nodes, requires an additional projection scheme. The projection to begin
a row with only side nodes is shown in Figure 14. Let the beginning (and ending) node by N,. Two
new nodes, Nj and Nk,are placed at the tips of vectors Vj and vk respectively. Vj originates at N ,
and vk originates at Ni+1. Vj is oriented to bisect the interior angle, a, at N,. vk is oriented to
bisect the interior angle, p, at N i + , .Let d , be the distance from Ni to Ni-,,
d , be the distance
from N, to N i + l ,and d3 be the distance from N i + l to Ni+1. The magnitudes of the vectors are
given by

Figure 14. Projection of new nodes from an all side node row

One new element, defined by nodes N , , N j , N , and N i + l ,is added to the mesh.
As can be seen in Figure 14, the projection from an all side node row generates two row end
nodes which serve to propagate and eventually terminate the forming row.

Row termination. With each of the projections described above, the node projected from them
becomes, in effect, a new row end node. Thus when the row has progressed such that the next
node, Ni,in the row is the row’s predetermined ending node, the row is terminated by simply
closing the final element in the row. The process of row termination is shown in Figure 15. The
final element in the row is defined by nodes N i P l ,N j , Niil and N , .

2.7. Smoothing
Smoothing is the most used tool during the paving process. It is used not only after rows are
added and intersections completed, as shown in the control sequence in Figure 4, but essentially
after every modification to the mesh. Local operations involving only a few nodes such as
projection during row generation, seaming, intersections, etc., tend to distort the neighbouring
mesh and paving boundaries. The smoother attempts to restore and maintain element size,
perpendicularity, and overall paving boundary and mesh smoothness; a set of criteria that are
often conflicting.
The smoothing requirements for paving differ from other mesh generation applications in that
a free boundary must be maintained as the paving boundary moves from the Jixed permanent
boundary toward the interior of the region. Thus, for paving we developed a customized three-
step smoother. First, all nodes not on the paving boundary are fixed and a smooth is applied to
any floating nodes on the paving boundary. This is called the paving boundary smooth. Second,
following the boundary smooth, all paving boundary nodes are fixed and all floating nodes, which
are not paving boundary nodes, are smoothed. This is called the interior smooth. Finally, the
boundary smooth is repeated. In this section we first describe the paving boundary smooth, next
the interior smooth, and finally a localization mechanism to reduce the cost of smoothing.

Figure 15. Row termination


Figure 16. Length adjustment to row nodes

Boundary node smoothing. As mentioned above, the boundary smooth is limited to nodes on
the current paving boundary that are not part of the original exterior boundary.$ The boundary
smoother is a modified isoparametric smooth.s4 To begin the smoothing process, the node
movement based on a true isoparametric smooth is calculated. Let Vi be the vector from the
origin to node Ni.Assuming that N iis attached to n elements, let Vmj, V,, and V,, be vectors
from the origin to respectively nodes Nj, Nk and N , of the mth element. The nodes must be in a
clockwise or counterclockwise order around the element. A new vector from the origin to the
proposed new location of node Ni,vector Vi, is given as

Let the vector A, define the change in location of node N , for a true isoparametric smooth. This
vector is then simply

The isoparametric smoother adjusts nodes such that the elements are more nearly para-
Ilelpipeds. However, it does not necessarily maintain the desired element size or the desired
squareness of an element. This becomes a problem for paving nodes that are attached to two
elements." For such paving nodes the change in node location is modified to reflect a desired
length and a desired angular smoothness.
First we will examine the desired length criteria. Let node Nj be the node opposite node N ion
the common boundary between the two elements attached to node Ni,as shown in Figure 16.
Again, let Vi and Vj be the vectors from the origin to node N , and Nj respectively. Let 1, be node

'Original exterior boundary nodes are never free to move

§These are usually boundary side nodes. As such, they constitute a majority of most boundary nodes during paving

N<s desired length, which is the length of the projection vector used to initially generate the node
as defined by equations (l),(2) and (5). Let 1, be the length of a vector V, from N j to the new
location of N i determined from the isoparametric smooth (i.e. to the tip of vector Vi in
equation (12)). A modified change in position vector, A,, is now defined as

This adjustment can be understood as simply a length adjustment of the vector V, without
adjusting its orientation, as shown in Figure 16. The length adjustment helps protect the desired
size of the element being generated as well as the aspect ratio. However, the perpendicularity of
element corners is not guaranteed.
The final adjustment during a paving row smooth is the desired angular smoothness. Again let
N j be the node opposite Ni on the common boundary, as shown in Figure 17. Let N,- and Ni+
be respectively the preceding and following nodes to N ion the boundary. Let the vectors Pi,Pi- ,
and Pi+ be vectors from node N j to nodes Ni, N i - and Ni+ respectively. A new vector, PB1, is
, ,
defined which bisects the angle between Pi- and Pi+ . Let a new vector, P,,, now be found. This
vector will determine the adjusted location of Ni. The tail of this vector is placed at N j . The
direction of the vector P,, is defined along the angle bisector of the angle between P,, and Pi.
Next the length of vector PB2 must be determined. First, the location of a new point, Q, is
calculated as the intersection between P,, extended, and a line connecting nodes Ni- and Ni+ . ,
Let the length from Q to N j be la and the original projection length of N ibe ID as before. The
length of vector PB2,I PB2 1, is then defined based on the relative lengths of ID and I,:

Figure 17. Angular smoothness adjustment to row nodes


The new location for N i based on angular smoothness has now been defined. The change in
position vector for angular smoothness, A,, is thus defined as
Ac = P,, - Pi (16)
This angle smoothness adjustment functions to keep elements perpendicular at corners and helps
to maintain a smooth paving boundary.
The final change in location vector, Ai, for node N i attached to only two elements on the
boundary can now be defined as

For nodes on the boundary which are attached to more or less than two elements, the change in
location is based solely on the isoparametric smoother vector change, A,, as defined in
equation (13). Although somewhat complex to implement, the row smooth is critical to the
correct formation of rows. It also converges quickly, usually in three or four iterations.

Interior node smoothing. Following the initial paving row smooth, all boundary nodes are fixed
and the remaining floating nodes are smoothed. A modified length-weighted Laplacian'
smoother is used. This technique uses a set of vectors from an interior node to all its attached
nodes. Let a contributing vector, C j , be the vector contribution of an attached node, N j , when
smoothing the interior node, N i . In defining C j , let V j be the vector from node N ito node N j , as
shown in Figure 18. If node N j is not on a permanent boundary, then the contributing vector C j is
equivalent to V j . If node N j is on a permanent boundary then C j is defined as
C j = V j + A,, (18)
where the vector A,, is defined using the angular smoothness criteria defined in equation (16).
Now, let the vector defining the change in location of N i be Ai. This vector, Ai, is defined as a
sum of all the contributing vectors from attached nodes weighted by the length of the

Figure 18. Modified length-weighted Laplacian smoothing applied to interior nodes


contributing vectors. If there are n nodes attached to N ias shown in Figure 18 then

A.I = j = 1

This modification to the length-weighted Laplacian tends to keep elements along the exterior
of the region perpendicular to the boundary. This smoother also tends to move nodes an equal
distance from its attached neighbours and converges rapidly.

Localization of smoothing. Since smoothing is an often repeated and fairly expensive proced-
ure, a localized smooth is used to minimize this cost. If a smooth mesh is changed through some
local operation it is reasonable to assume that, during the subsequent smooth, nodes closer to the
disturbance will require more adjustment than those more distant. In fact, nodes buried more
than three element layers from the disturbance rarely move significantly. Thus, in order to save
computer time, smoothing has been restricted to nodes connected within three element layers to
any node involved in a disturbance. For instance, when a row is added, only those nodes three
rows deep into the existing mesh are smoothed. This simple localization has improved run times

2.8. Seaming
The process of seaming may be understood as simply a process to close cracks in the paving
boundary which often form during the paving process. Crack tips are easily recognized as
boundary nodes with very small interior ang1es.v
The criterion to seam a crack tip is based on the interior angle of the node, a, the number of
attached element edges, NE,and whether the node is a fixed node (i.e. on a permanent boundary).
There are three types of seams which must be considered: (i) a seam of a floating paving boundary
node, (ii) a seam of a fixed node, and (iii) a seam between two element sides of unequal length (a
transition seam).

Interior node seams. The seaming of floating nodes on a paving boundary is the most common
type of seam. Such nodes may be seamed if
a <el for NE 2 5
where < c2
a < E~ otherwise
As seen in equation (20) above, the seaming tolerance decreases as the number of attached
elements increase. This tolerance shift favours seams which will form a regular node or reduce the
irregularity of the node at the crack tip. For example, the node, Ni, in Figure 19(a)is attached to
four element edges (NE= 4). Seaming this node, as shown in Figure 19(b), changes N ito an
irregular node which is attached only to three elements. Thus, in general, the interior angle of a
node attached to four element edges must be relatively small for this type of seam to be beneficial.
In this case the angle at node N i was negative and seaming was necessary. On the other hand, a
seam of node Ni, shown in Figure 19(c),which is attached to five element edges, changes Ni into a
regular interior node attached to four elements as shown in Figure 19(d).This type of seam is thus
allowable for much larger interior angles.
$It is not unusual for rows to cross at their ends, thus forming nodes with negative interior angles

a) 4 - L i n e Crack Node b l 4-Line Node Seamed

cl 5-Line Crock Node dl 5-Line Node Seamed

Figure 19. Seaming floating node cracks in the paving boundary

As can be seen in Figure 19, each seam is performed by simply combining the node preceding
Ni,N i - 1 , and the node following N , , N i + ,into a single node, N j . This also reduces the number
of element edges, N , , attached to node N i by one. The example seam shown in Figure 19 also
shows that once a crack begins to close (19(b))the seam tends to continue until a fairly square
corner is formed (19(d)).

Fixed node seams. The seaming of cracks involving fixed nodes on the permanent boundary
must be performed more judiciously. Since exterior nodes are not free to move, smoothing cannot
improve elements which have two or more sides on the exterior boundary. Thus, care must be
taken that a seam will not create elements with bad interior angles which cannot be improved.
This is primarily a concern when the exterior boundary forms small angles. As shown below, the
problem is overcome by simply delaying the closure of cracks where this becomes a problem.
Figure 2qa) shows an example of the fixed node seam problem. The first row placed into the
small angle portion of the permanent boundary forms a very small interior angle at node N , .
However, seaming node Ni would create a degenerate quadrilateral element with an interior
angle of % 7c. As shown in Figure 20(b), the seam of this crack is thus ignored and the next row
inserted as if the small crack did not exist. This, of course, causes the new row to seriously overlap.
But now the nodes to be seamed are all floating nodes and can be smoothed as needed. The mesh
resulting from this delayed seam is shown in Figure 2qc). For geometries such as this which
contain small interior angle portions, the quadrilateral mesh generated by the paving algorithm is
almost always similar to the mesh shown in Figure 2qc).

0 1 First Row I n s e r t e d b l Second R o w I n s e r t e d

cl Seamed Corner
Figure 20. Seaming fixed node cracks in the paving boundary

a ) Mesh To Be Seamed bl Wedge I n s e r t e d B e f o r e Seam

cl Seam Performed d l Mesh Smoothed

Figure 21. Seaming disproportionateelements in the paving boundary

Transition seams. Often, for paving boundaries with transitioning requirements, a seam may be
required between two element sides of disproportionate sizes, as shown in Figure 21(a). Thus,
before any seam is performed, the relative lengths of the two element sides to be seamed are
compared. If the ratio of the longer side to the shorter side is greater than a threshold value,
currently set at 2.5, then a wedge element, as will be explained in Section 2.9, is added to the mesh.
The wedge is positioned so as to effectively cut the longer element side into thirds, as shown in
Figure 21(b)." This new arrangement of elements is then seamed, as shown in Figure 21(c), and the

"Thewedge or triangular looking element in Figure 21(b) is actually a quadrilateral with one of the nodes at the midpoint
of the triangle's top side

result of such a seam after smoothing is shown in Figure 21(d). As can be seen in this example,
insertion of a wedge during the seaming operation allows graceful transitions in otherwise
difficult boundary closure situations.

2.9. Row adjustment

As rows are added to the mesh, the geometry, as well as the topology of the row, affect the
elements being formed. For row portions that traverse a concave segment of the boundary, the
elements in each successive row tend to become smaller. Likewise, if the row portions traverse a
convex segment of the boundary, elements tend to become larger with each additional row. This
undesired expansion or contraction of element sizes, depicted in Figure 22, can be corrected with
appropriate row adjustments. This section describes the techniques used to recognize and correct
these phenomena: first the insertion of wedges in expanding rows, and second the formation of
tucks in contracting rows.

Insertion of wedges. Wedges are used to check the growth of the size of elements in portions of
a boundary row. The wedge insertion process can be visualized as cracking the row at a floating
paving boundary node and forcing an additional element into the crack. This new element
essentially forms a wedge which separates two previously connected elements and forms a row
corner node in the row.
In practice wedge insertion entails the generation of two new nodes and one new element.
Referring to Figure 23(a), for the insertion of a wedge at node N , , a new node, N I, is generated
and used to replace N , in the element definition of the element on the right of N i . Node N , is
repositioned on an imaginary line joining node M i - and the original position of N iat a distance
of 1/3 the original distance from Ni to N i - 1 . Likewise, node Nf is positioned on an imaginary line
joining the original position of N , and node Ni+ at a distance of 1/3 the original distance from N ,
to Ni+1 . This forms the crack referred to earlier. The second new node, N k , is then placed at a

Figure 22. An example of a row with portions of contracting and expanding element sizes

(a1 O r i g i n a l Mesh [ b l Wedge I n s e r t i o n (cl Smoothed Mesh

Figure 23. The insertion of a wedge to correct element size expansion

(a1 Boundary Before Wedges (bl Boundary A f t e r Wedges

Figure 24. Example wedge insertion in a boundary row portion

location defined using the isoparametric smoothing equation, equation (12), using nodes N , , N j
and N in the equation. The new element is then defined by nodes N , , N j , N : and N , , as shown in
Figure 23(b). Smoothing is then performed and the results are shown in Figure 23(c).
Row portions where wedge insertions would be beneficial are detected based on two criteria:
first, on an expansion ratio which is the ratio of the desired length* to the actual distance between
adjacent paving boundary nodes, and second, on the interior angle of the node. If the expansion
ratio and interior angle are both greater than threshold values,+ then the node qualifies for the
insertion of a wedge. The actual number of wedges and their location is based on the number of
sequential paving boundary nodes that qualify for wedge insertion, and the total angle traversed
by the boundary for this sequential set of nodes. The optimum wedge placement would be at the
midpoint of every n/2 portion of the total angle traversed by this sequential set of paving
boundary nodes. For example, in Figure 24(a), all paving boundary nodes in the interior paving
boundary (around the interior hole) qualify for wedge insertion. Since the total angle traversed by
this sequence of eligible nodes is 2n, and since there are more than four nodes in the sequence, four

*The desired length is usually propagated from the boundary inward and originates as the distance between nodes along
the permanent boundary
'The expansion ratio threshold is currently set at 1.25 while the interior angle threshold is 183"

(a) OriginaL (bl E l e m e n t (cl S i d e s [ d l Smoothed

Mesh De Le t e d S u p e r imposed Mesh
Figure 25. The formation of a tuck to correct element size reduction

wedges are inserted at nodes NA, N , , N , and N,, as shown in Figure 23(b). The nodes where
wedges are placed are located at the clockwise 118,318,518and 718 points (midpoints of each x/2
portion) of the traversed angle. No attempt is made to insure alignment of these wedges with
other existing boundaries, although in this example alignment does occur. After all of the wedges
have been inserted into the row portion, a localized smooth is performed.

Formation of tucks. Tucks are used to check the reduction of the size of elements in portions of
a boundary row. The tuck formation process can be visualized as combining two adjacent row
elements into one. This combination process usually creates a row end node where three row side
nodes previously existed.
A tuck formation eliminates two boundary nodes and one element. Referring to Figure 25, for
the formation of a tuck at node Ni,an element must first be deleted. Let element side SA be the
element side connecting nodes N , and N j . The element attached to side S, is the one deleted, as
shown in Figure 25(b). Let element side S, connect nodes N , and N j , element side SL connect
nodes N , and N , , element side S , connect nodes N j and N j + 1 , and element side S&connect N i and
N , . The tuck is completed by superimposing sides S, and S , on sides SL and S&respectively, as
shown in Figure 25(c). This overlay replaces node N j with node N , and node N j + with node N i
within all associated elements. A smooth completes the tuck and produces the mesh shown in
Figure 25(d).
As with wedges, row portions where tuck formations would be beneficial are detected based on
two criteria: first, on a reduction ratio which is the ratio of the desired length to the actual
distance between adjacent boundary nodes, and second, on the interior angle of the node. If the
reduction ratio and interior angle are both less than threshold values,’ then the node qualijes for
the formation of a tuck. Again, as with wedges, the actual number of tucks and their location is
based on the number of sequential boundary nodes that qualify for tuck formation and the total
angle traversed by the boundary for this sequential set of nodes. The optimum tuck placement
would be at the midpoint of every x / 2 portion of the total angle traversed. For example, in
Figure 26(a), three sequential sets of nodes exist which qualify for tuck formation. The first set is
the top right boundary node N , , the second set is the left loop counterclockwise from node N 2 to
N , , and the third set is the lower right boundary node N , . Of course, since the first and third set
are composed of only one node, only one tuck is taken at the node in each set. The angle traversed
by the second sequential set is x 3x12 and there are more than three nodes in the sequence. Thus

*The reduction ratio threshold is currently set at 0 8 while the interior angle threshold is 177”

three tucks are formed at nodes NA, N , and N,, as shown in Figure 23(b). Nodes NA, N , and N ,
are at the counterclockwise 116, 112 and 516 points (midpoints of each % z/2 portion) of the
traversed angle, respectively.

2.10. Perimeter intersection handling

As paving boundaries propagate into the region, it is common to have these boundaries
intersect or overlap existing portions of the mesh. These intersections may be associated with a
negative interior angle and thus be completely eliminated during the seaming operation, as
previously shown in Figure 19. However, many times the overlap is simply due to the original
geometry. As shown in Figure 27, the overlap may be of two types: (i) a single paving boundary
crossing itself, and (ii) a paving boundary crossing another paving boundary (e.g. holes within a
region). This section describes the mechanism for detecting and correcting these two types of

Single paving boundary intersection. When a paving boundary overlaps itself, as shown in
Figure 27(a), there are at least two pairs of lines (connections between boundary nodes) that

a1 Boundary B e f o r e Tucks bl Boundary A f t e r Tucks

Figure 26. Example tuck formation in a boundary row portion

a1 S i n g l e Boundory
bl M u l t i p L e Boundories
Figure 27. Examples of perimeter intersections requiring adjustment

Figure 28. Intersection calculation vectors

intersect each other. The intersection is detected by calculating the intersection of two straight
line segments (Reference 16, pages 319,320). The intersection calculation formula used is fast and
informative, providing not only the verification of an intersection but also the location of the
intersection and the proportional lengths of the four line segments that would be formed if the
lines were divided at the intersection. As shown in Figure 28, let the vector V, be the position
vector from the origin to end point A of line segment a.
Let vector V, be a vector from end
point A to end point B of line segment a.
Likewise, let V, be the position vector of end point C
of line segment CD and VD be a vector from C to D of line segment z.
A position vector P to
any point along line segment may then be described as P(u) = A + uB, where U E [0,1].
Similarly, a position vector Q to any point along second line CD may be described as Q(w) = C
+ wD, where w E [0,1]. Let R be the position vector of the point of intersection. If an intersection
exists then
R = P(u)= Q(w) (21)
A + uB = C + wD
In two dimensions, this leads to two linear equations (one for each co-ordinate direction) and
two unknowns, u and w:
A, + uB, = C , + wD, (23)
A, + uB, = C , + wD, (24)
In the solution of these equations, if 0 < u < 1, then the lines intersecteand the relative location
of the intersection is indicated by the values of u and w. For instance, if u = 0.25, then the
#If 0 < u 6 1 then the relation 0 Q w < 1 must also hold

intersection occurs at the quarter point along line AB

closest to point A. A similar relation holds
for the value of w and line m. This type of information is essential in properly correcting the
overlap, as explained below.
On& an intersection has been detected, it is corrected by connecting the overlapping portions
in the best possible manner. For this case of a single boundary intersecting itself, the connection
will result in two new paving boundaries being formed. In many cases the best connection will be
between the two lines that intersect. However, it is often more advantageous to combine lines
adjacent to the intersection lines on either side and in either direction. The following considera-
tions are followed to insure that the connection formed will be valid and will avoid the formation
of irregular nodes where possible:
0 Evenness constraint. The two new paving boundaries must each contain an even number of
elements, as explained in Section 2.1. If connecting the intersecting lines will produce paving
boundaries with an odd number of nodes, the connection location is adjusted based on the
relative values of u and w.
0 Forward/backwardsearch. In some cases, connecting the intersecting lines would produce a
more irregular mesh than connecting the two lines either forward or backward one node of
the actual intersection. Thus connecting lines one step in either direction are compared based
on nearness and how close the lines are to being parallel.
0 Size diflerences. If the length of the respective intersectinglines differ significantly,a wedge is
inserted similar to the process described in Section 2.8 to enable an easier transition.
0 Explore all intersections. There will always be an even number of intersections of a closed
paving boundary upon itself. By connecting one intersection and then seaming, as explained
below, at least one more intersection location is eliminated. Thus, all intersections are found
and compared and the best one (again based on the nearness and parallelness of the two
intersecting lines) connected first.
Once the most appropriate connection has been found and the lines combined, a seam is
attempted for both of the newly formed boundaries. Since any boundary portions which remain
overlapped will form negative internal angles at the point of connection, a seam will pull all of
these back into the existing mesh. Since multiple intersections may occur with the insertion of any
row, each of the newly formed paving boundaries are themselves checked for remaining
intersections before proceeding.
Figure 29 illustrates the connection of an intersecting boundary. In Figure 29(a), a new row has
been added which overlaps portions of the existing paving boundary. In this case there are two
intersection locations found, as highlighted by the dashed lines in Figure 29(a). The bottom
intersection pair could be connected to form two loops with an even number of boundary nodes.
However, the top intersection pair would form two loops with odd numbers of boundary nodes,
and so steps in either direction are considered. The match highlighted in Figure 29(b) is the best
connection for the top intersection. This connection is also preferable to the bottom connection
based on nearness to being parallel and is thus connected as shown in Figure29(c). In this
example the seam operation following the connection closes both of the new paving boundaries
formed and the completed mesh is shown in Figure 29(d).

Multiple boundary intersections. When a paving region contains interior paving boundaries
(holes), they must all be eventually connected before closure is possible. Intersections are found
similar to the intersection checking performed for a single boundary intersecting itself. However,
since each boundary must contain an even number of intervals, it is not possible to generate an
odd boundary loop from the connection of any two paving boundaries. Thus, connection is based

only on the goodness of the connection. This type of intersection may occur between the external
and an internal paving boundary or between any two internal paving boundaries.
Figure 30 illustrates the intersection and connection of multiple boundaries. As shown in
Figure 30(a), the exterior paving boundary intersects three of the interior paving boundaries
(circular holes). The intersections with these three boundaries are evaluated, the most appropriate

a1 I n t e r s e c t i o n s Found b 1 Connection Chosen

cl Connection Formed d1 Seams Performed

Figure 29. Example single boundary intersection formation

a) Row I n s t a l l e d b ) 3HoLesConnected

c 1 One Boundary d l CompLeted Mesh

Figure 30. Example multiple boundary intersection formation

connection chosen, connected and seamed as shown in Figure 3qb). At this point there are only
two paving boundaries left, the exterior paving boundary and the interior boundary around the
centre hole. These two boundaries are then paved until they overlap and are connected as shown
in Figure 30(c), forming a single exterior boundary. The completed mesh is shown in Figure 30(d).

2.11. Cleanup of completed mesh

After the region has been completely paved, local adjustments are attempted to improve the
overall quality of the finished mesh. These adjustments are essentially an insertion or deletion of
an element to improve the local aspect ratios, reduce the number of irregular nodes, or eliminate
elements with bad interior angles. Factors controlling the adjustments include the number of
element edges attached to a node, the interior angle of all elements attached to the node and the
anticipated probability of improving the mesh locally.
Figure 31 shows an example of an element deletion improving the mesh. In Figure 31(a), the
element formed by nodes N , , N j , Nk and N , has been targeted for deletion. Of the nodes which
form this element, only N , is a regular node. Nodes N j and N , are attached to three elements and
node N k is attached to five elements. The deletion is performed by collapsing the element as
shown by the arrows in Figure31(a). Nodes N j and N , and the attached element edges are
combined during the deletion. As a result of the deletion one regular node, N,, is transformed into
an irregular node, as shown in Figure 31(b). However, two irregular nodes, N j and N , , are

a1 Etement To Be D e l e t e d bl A d j u s t e d Mesh

Figure 31. Example of an element deletion adjustment

a) Location For I n s e r t i o n b l R d j u s t e d Mesh

Figure 32. Example of an element insertion adjustment

combined into one regular node, and another irregular node, N,, is transformed into a regular
node. Thus, the overall quality of mesh has been improved.
Figure 32 shows an example of an element insertion to improve the mesh. In Figure 32(a), the
two elements highlighted with dashed lines have been targeted for an element insertion. These
type of elongated elements usually only occur during paving when a large transition is necessary.
Figure 32(b) shows how the insertion of an element not only decreases the aspect ratios of these
elements, but also improves the interior angles of the elements.

In order to visualize how each of the steps of paving interact during the meshing process, several
example problems are shown in this section. These problems are shown in a quasi-animation
sequence. Each frame in the figure represents the mesh at a stage in the progression of the
algorithm beginning at the upper left in each figure and progressing left to right for each row top
to bottom. A quantitative evaluation of the quality of the mesh for each of the example problems
is included in Section 4.

F'igure 33. Paving of Example P


3.1. Example A
Figure 33 depicts the paving process for Example A. Example A consists of a square exterior
permanent boundary with a square interior permanent boundary which is skewed with respect to
the exterior boundary.
Paving proceeds alternately around the exterior and the interior paving boundary until the two
overlap, as shown in Figure 33(c). The interior boundary is then connected to the exterior in
Figure 33(d). In this case, two exterior paving boundaries are formed. One of these is completely
paved before the other is continued as shown in Figure 33(e). This process continues until all the
exterior boundaries close, resulting in the mesh shown in Figure 33(h). The mesh after invoking
cleanup procedures is shown in Figure 33(i). In this case cleanup does not affect the resulting
This example shows the boundary sensitivity of the paving process. Irregular nodes have been
pushed well away from permanent boundaries and the mesh along all permanent boundaries is
well formed.

3.2. Example B
Figure 34 depicts the paving process for Example B. Example B consists of an exterior
boundary with two interior slots.
The exterior permanent boundary contains a non-uniform spacing of nodes. Nodes on the
concave arc are approximately twice as dense as on other portions of the boundary. Likewise, the
density of nodes for the upper slot is about that of the concave arc on the exterior boundary, while
node density on the lower slot is more nearly that of the rest of the exterior boundary.
This example presents some interesting aspects of paving. Figure 34(b) shows the overlap of
two paving rows of quite dissimilar element sizes. However, the seaming algorithm performs
nicely to connect these rows with a minimum of irregular nodes, as shown in Figure 34(c). The
frames shown in Figure 34(d), (e) and (f) show the effects of row adjustment. Adding the second
paving row around the lower slot produces quite large elements, as shown in Figure 34(d). Four
element wedges reduce element sizes appropriately and allow subsequent connection of this
interior paving boundary to the exterior paving boundary, as shown in Figure 34(e) and (f). Final
closure is shown in Figure 34(g), (h) and (i). Qualitatively, the final mesh is well formed and
boundary sensitive in spite of the mesh gradation requirements of this example.

3.3. Example C
Figure 35 depicts the paving process for Example C . Example C consists of a very odd shaped
exterior permanent boundary with one interior permanent boundary. The exterior boundary
contains two sharp corners, one on the left centre of the geometry and the other on the right lower
portion. The interior boundary is also close to the exterior in terms of the number of element rows
which can be inserted between the two boundaries.
This example shows the effect of sharp angles in the original geometry as well as the need for
the cleanup algorithm. In Figure 35(b), one element layer has been added to all boundaries. The
first exterior row has produced a seam in each of the two sharp corners. This seam blunts the
corner, allowing for a more regular mesh to be inserted in subsequent steps. The severe overlap
between the exterior and interior paving boundary is also shown in Figure 35(b). Elements
become appreciably distorted when this overlap is corrected as shown in Figure 35(c). This
distortion remains in the mesh through the final closure shown in Figure 35(h). However, the
Figure 34. Paving of Example B Figure 35. Paving of Example C

cleanup technique is able to eliminate most of the distorted elements, as shown in the final mesh in
Figure 35(i).

3.4. Example D
Figure 36 depicts the paving process for Example D. Example D consists of a simple exterior
permanent boundary with no interior boundaries. The exterior boundary does have an extreme
gradation in the density of nodes along the boundary. As can be seen in Figure 36, nodes on the
boundary of the notch at the upper right of the geometry are on the order of twenty times more
dense than the nodes along the left and bottom boundary segments. The top and right boundary
segments have an increasing density of nodes as they approach the notch portion.

Figure 36. Paving of Example D


This example shows the versatility of the paving algorithm to handle large mesh gradation
requirements. Paving rows retain the size dictated by the boundary segments from which they
originate, as shown in Figure 36(d), (e) and (f). As disproportionate rows interact with each other
adjustments are made to insure validity of the mesh. The mesh after final closure is shown in
Figure36(h). Cleanup dramatically improves the quality of the final mesh, as shown in
Figure 36(i). As shown in this example, transitions tend to be tight or close to the areas of smaller
mesh density. This is because a large element row, by definition, moves farther from the
permanent boundary than a small element row. Thus when intersections occur between small and
large rows, the intersection is physically much closer to the permanent boundary of the small

As mentioned earlier, the goal of this research is to generate an all-quadrilateral mesh in an
arbitrary geometry. The criteria used to evaluate the performance of paving include (i) the quality
of the generated mesh, (ii) the robustness of the algorithm, and (iii) the speed of generation. The
example problems shown in Section 3 present a qualitative view of the paving algorithm. In this
section we attempt to quantify these results.

4.1. Mesh quality

The qualitative objectives listed in Section 1 have been met by this technique. As can be seen in
the examples, the meshes generated are boundary sensitive, orientation insensitive and minimize
the placement of irregular nodes.
However, this type of qualitative assessment of the technique is of limited use in the very
quantitative field of finite element analysis. In an attempt to quantify the quality of a mesh
produced with the paving technique, an element distortion metric, proposed by Oddy et al.” for
isoparametric elements, has been used. This distortion metric measures the degree of deviation of
an element in the physical domain (x,y , z ) from its assumed shape in the computational domain
(5, q, It combines effects of shearing and stretching, is not affected by rigid body motions and is
independent of element size. In calculating this metric we assumed a 4 node linear-order
isoparametric element with single point integration.
As stated in Reference 17, a distortion metric can be used only as a relative indication of error
since factors other than the mesh (e.g. type of analysis, size of elements and loading) affect analysis
accuracy. However, elements with distortion metric values greater than 5 tend to be problematic.
As an aid in understanding this distortion metric, a surface plot of the metric values for various
parallelograms is shown in Figure 37. In this figure the aspect ratio is the ratio of lengths of the
non-parallel sides of the parallelogram and 0 is the acute angle between them.
Let D be the value of the distortion metric for an element. Average values of D, D, for meshes
generated using the paving technique range from 0 0 for blocky type geometries to 0.1 for
relatively complex geometries without transitioning requirements. Transitioning, of course,
requires the use of more distorted elements, but values of D still tend to be around 0.3. These
results support the qualitative observation that paving tends to produce meshes with square
As examples of the distortion metric evaluation, the meshes generated in Section 3 have been
used. The element with the maximum D value has been indicated with the asterisk (*) symbol in
Figures 38-41. Figure 38 shows the mesh generated for Example A. In this mesh the average
element distortion metric, D, is 0-02,with a maximum D value of 0.26. Figure 39 shows the mesh

Figure 37. Surface plot of distortion metric for parallelograms

1 1 1 1 1 1 1 1 1 ~ 1 1 I 1 I ~ I ~1 1 1

Figure 38. Example A: d = 002, D,,, = 0.26

for Example B. In this mesh d is 0.09, with a maximum value of D of 1-09.Figure 40 is the mesh
for Example C. Here d is 0.1 1and the maximum D value is 1.13. The final example, Example D, is
shown in Figure 41. For this example d is slightly higher at 0-30 with the maximum D value of
1-96.This value is for the element in the upper left corner where the intervals on the exterior
boundary force a 3: 1 aspect ratio at the corner.

Figure 39. Example B: D = 0.09,D,,, = 1.09

Figure 40. Example C: d = 0.11, I),,, = 1.13

4.2. Robustness
As mentioned earlier, paving has proved empirically to be robust for very complex geometries.
Even drastic transitions have been handled well with paving. However, the technique breaks
down when impossible transitions (e.g. 30: 1) are required over a very restricted area. However,
even this limitation can be removed if adjustment of the permanent boundary (i.e. insertion/dele-
tion of nodes) is allowed.

4.3. Speed
The speed of the paving approach is difficult to tie directly to the size or number of nodes in a
generated mesh. This is because paving’s effectivenessis tied to the shape of the boundary as well
as the number of nodes in the boundary. For instance, the paving boundary for a geometry which

Figure 41. Example D: d = 0.30, D,,, = 1.96

flesh Nodes

Figure 42. Required CPU for paving vs. number of nodes in the mesh

is relatively long and thin tends to overlap itself quickly, generating many smaller paving
boundaries which can be processed rapidly. On the other hand, large open boundaries are
processed somewhat more slowly. We have measured performance empirically and have found
that the required CPU time is roughly linear with the number of nodes in the final mesh, as shown
in Figure 42. The geometric dependencies account for the scatter in the data. For this study, a
relatively unloaded VAX 11/780 computer was used to process various geometries with increas-
ingly finer mesh requirements. In practice, most meshes are being generated quickly enough to
allow interactive processing.

The technique of paving has satisfied the objectives of automatically generating a high quality
mesh for an arbitrary geometry. The technique has been empirically shown to be robust and fast
for complex geometries and extreme mesh gradations. The paving technique tends to place well
formed elements along the boundary with irregular nodes in the interior of the geometry. The
mesh quality measured using a distortion metric has been shown to be extremely good. This
technique is thus suitable for general automated all-quadrilateral mesh generation.
This technique may prove very powerful when combined with an error estimator for iterative
adaptive analysis. The use of a function defining the required element size over the entire
geometry rather than only at the boundaries should improve the technique. The use of an
approach similar to paving for 3-D all-hexahedral mesh generation is also being investigated.

1. P. L. Baehmann, S. L. Wittchen, M. S. Shepard, K. R. Grice and M. A. Yerry, ‘Robust geometrically based, automatic
two-dimensional mesh generation’, Int. j. numer. methods eng., 24, 1043-1078 (1987).
2. A. Kela, R. Perruchio and H. B. Voelker, ‘Toward automatic finite element analysis’, Camp. Mech. Eng., 5, 1 (1986).
3. M. S. Shepard and M. A. Yerry, ‘Approaching the automatic generation of finite element meshes’, Comp. Mech. Eng.,
1. 49-56 (1983).
\ I

4. B. Joe and R. B. Simpson, ‘Triangular meshes for regions of complicated shape’, Int. j. numer. methods eng., 23,
751-778 (1986).
5. B. Wordenweber, ‘Finite-element analysis for the naive user’, in M. S. Pickett and J. W. Boyse (eds.), Solid Modeling by
Computers, Plenum Press, New York, 1984, pp. 81-102.
6. A. J. G. Schoofs, L. H. T. M. Van Beukering and M. L. C. Sluiter, ‘A general purpose two-dimensional mesh
generator’, Aduan. Eng. Software, 1-3, 131-136 (1979).
7. W. A. Cook and W. R. Oakes, ‘Mapping methods for generating three-dimensional meshes’, Comp. Mech. Eng., 1,
67-72 (1983).
8. W. J. Gordon and C. A. Hall, ‘Construction of curvilinear co-ordinate systems and applications to mesh generation’,
Int. j . numer. methods eng., 7, 461-477 (1973).
9. K. Ho-Le, ‘Finite element mesh generation methods: A review and classification’, Comp. Aided Des., 20,27-38 (1988).
10. M. S. Shepard, ‘Finite-element mesh generation for use with solid modeling and adaptive analysis’, in M. S. Pickett
and J. W. Boyse (eds.), Solid Modeling by Computers, Plenum Press, New York, 1984, pp. 53-80.
11. T. D. Blacker, M. B. Stephenson, J. L. Mitchiner, L. R. Phillips and Y. T. Lin, ‘Automated quadrilateral mesh
generation: A knowledge system approach‘, ASME Paper No. 88- FVA/ClE-4, 1988.
12. T. D. Blacker, FASTQ Users Manual Version 1.2, SAND88-1326, Sandia National Laboratories, 1988.
13. M. B. Stephenson and T. D. Blacker, ‘Using conjoint meshing primitives to generate quadrilateral and hexahedral
elements in irregular regions’, Proc. ASME Computations in Engineering Conference, 1989.
14. L. R. Herrmann, ‘Laplacian-isoparametric grid generation scheme’, J . Eng. Mech. Diu. A X E , 102, EMS, 749-756
15. R. E. Jones, QMESH: A self-organizing mesh generation program, SLA-73-1088, Sandia National Laboratories, 1979.
16. M. E. Mortenson, Geometric Modelling, Wiley, New York, 1985.
17. A. Oddy, J. Goldak, M. McDill and M. Bibby, ‘A distortion metric for isoparametric finite elements’, Trans. CSME,
No. 38-CSME-32, Accession No. 2161 (1988).